An EMD-based chaotic least squares support vector machine hybrid model for annual runoffforecasting

Abstract

Accurate forecasting of annual runoffis necessary for water resources management. However, a runoffseries consists of complex nonlinear and non-stationary characteristics, which makes forecasting difficult. To contribute towards improved prediction accuracy, a novel hybrid model based on the empirical mode decomposition (EMD) for annual runoffforecasting is proposed and applied in this paper. Firstly, the original annual runoffseries is decomposed into a limited number of intrinsic mode functions (IMFs) and one trend term based on the EMD, which makes the series stationary. Secondly, it will be forecasted by a least squares support vector machine (LSSVM) when the IMF component possesses chaotic characteristics, and simulated by a polynomial method when it does not. In addition, the reserved trend term is predicted by a Gray Model. Finally, the ensemble forecast for the original runoffseries is formulated by combining the prediction results of the modeled IMFs and the trend term. Qualified rate (QR), root mean square errors (RMSE), mean absolute relative errors (MARE), and mean absolute errors (MAE) are used as the comparison criteria. The results reveal that the EMD-based chaotic LSSVM (EMD-CLSSVM) hybrid model is a superior alternative to the CLSSVM hybrid model for forecasting annual runoffat Shangjingyou station, reducing the RMSE, MARE, and MAE by 39%, 28.6%, and 25.6%, respectively. To further illustrate the stability and representativeness of the EMD-CLSSVM hybrid model, runoffdata at three additional sites, Zhaishang, Fenhe reservoir, and Lancun stations, were applied to verify the model. The results show that the EMD-CLSSVM hybrid model proved its applicability with high prediction precision. This approach may be used in similar hydrological conditions.